CropHealthyNet: A Lightweight Hybrid Network for Efficient Crop Disease Detection

Abstract

Deploying high-precision deep learning models on resource-constrained edge devices remains a challenge for agricultural disease detection. This study introduces CropHealthyNet, a lightweight hybrid architecture optimized for both accuracy and computational efficiency. The architecture incorporates three key components: the ExGhostConv module, which integrates FReLU and SimAM attention for enhanced feature utilization; a Universal Position Encoding mechanism that adaptively captures spatial information to address variable lesion scales; and a MemoryEfficientTransformer employing chunked attention to mitigate global modeling memory overhead. Experiments on CDC, AGD_256, and CornLeafDisease datasets indicate that CropHealthyNet achieves a weighted average accuracy of 90.55% with 0.47 million parameters. The model outperforms several state-of-the-art lightweight architectures and achieves accuracy comparable to DenseNet121, with approximately 15 times fewer parameters. These results position CropHealthyNet as a viable solution for real-world deployment in resource-limited agricultural environments.

Plant phenotyping relevance

植物病害（病斑）を画像から検出する深層学習モデルの開発・評価が研究の中心であり、植物の病害状態を推定するフェノタイピング手法に該当する。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.